Patent Pending

Agent Lifecycle Governance

A unified system for governing autonomous AI agents from registration through active operation to retirement. Measuring and controlling the total potential impact of every agent in real time.

Patent Application

A Unified System for Design-Time and Real-Time Governance of Autonomous Software Agents

A system and method for governing autonomous software agents addresses the technical problem of cumulative operational authority. The system features a design-time Agent Registry as an authoritative system of record. This registry issues a verifiable cryptographic identity and stores a multi-layered data model defining the agent's approved mission. A risk quantification engine uses this model to programmatically calculate the agent's maximum potential impact and its excess, un-needed authority prior to deployment. This design-time plane is synergistically coupled with a real-time enforcement plane. The real-time plane validates the agent's identity and uses the authoritative profile and its pre-calculated risk quantifications from the Registry as non-bypassable constraints for making just-in-time, per-action privilege decisions. This unified architecture enforces least-privilege across the entire agent lifecycle.


Filed: November 2025 (Singapore) | Status: Patent pending

The Problem

When you give an AI agent the ability to act autonomously, how do you control what it can actually do? Traditional access management gives users fixed permissions. But AI agents are fundamentally different.

Agents inherit permissions from whoever calls them. They chain calls to other agents, each of which may inherit permissions from the first. They operate in environments they were not originally designed for. They drift from their declared purpose without any explicit permission change. The cumulative effect is that an agent's real authority at any moment may far exceed what was explicitly granted.

This "cumulative operational authority" is invisible to conventional security tools. A standard IAM system can tell you what permissions were granted to an agent. It cannot tell you what the agent can actually do when you factor in delegation chains, inherited permissions, and accumulated access across multiple systems. And it certainly cannot tell you what the total potential damage would be if that agent were compromised.

How It Works

The system couples a design-time registry (the authoritative record of what an agent should be allowed to do) with a real-time enforcement engine (the system that decides what it is allowed to do right now, for this specific action).

Digital identity

Every agent gets a verifiable digital identity, a "Digital Birth Certificate" based on industry-standard cryptographic protocols. This identity is coupled with a structured profile describing the agent's mission, boundaries, permissions, and capabilities. The profile is organised into ten layers covering identity, authority, delegation chains, data access, environment context, tools, capabilities, policy rules, risk metrics, and declared purpose.

Design-time registry

The Agent Registry serves as the single source of truth for every agent in the organisation. Before an agent is deployed, its profile is reviewed and approved by a designated steward. The registry supports multiple entry paths: manual registration by developers, automated discovery of running agents in live environments, and detection during the build pipeline. Regardless of how an agent enters the registry, it must have an approved profile before it operates.

Real-time, per-action decisions

At runtime, each time an agent requests access to a resource or attempts an action, the system makes a just-in-time privilege decision. It retrieves the agent's approved profile from the registry, assembles a real-time contextual graph incorporating delegation chains, target resources, and current environment conditions, and computes a dynamic risk score. Based on this score and versioned policy thresholds, the system grants, denies, downgrades, or sandboxes the request. Every decision is accompanied by a human-readable explanation citing the specific policy rules and risk factors involved.

Ephemeral credentials

When access is granted, the system issues time-limited credentials scoped to the specific action. These credentials are automatically revoked when the task completes, when their time window expires, when policy changes, or when the system detects anomalous behaviour. This replaces the traditional model of persistent, broadly scoped permissions with minimal, time-bounded access that expires by default.

Blast Radius

"Blast Radius" metrics measure the total potential impact of each agent: what could go wrong if this agent were compromised or behaved unexpectedly? The system calculates multiple variants, including a static design-time estimate, a dynamic runtime calculation, and a simulated future-state projection for pre-deployment impact assessment. For teams of agents working together, an aggregated Blast Radius captures the combined exposure.

Development pipeline integration

The system integrates directly into CI/CD pipelines. When a deployment is initiated, a governance gate intercepts the change, compares the proposed agent against its approved profile, recalculates risk metrics, and blocks non-compliant deployments. A simulation engine allows developers to perform "what-if" analysis before committing code, projecting the governance impact of proposed changes.

What Makes It New

Existing IAM systems grant static permissions and assume human operators. The Unified Governance architecture couples design-time governance (what should agents be allowed to do?) with real-time enforcement (what should this agent be allowed to do right now?), treating cumulative operational authority as a measurable, governable quantity.

linkDesign-Time and Runtime Coupled

The approved profile feeds runtime enforcement. Runtime telemetry feeds back into the registry through reconciliation loops. The two planes are connected, not independent.

calculateAuthority as a Measurable Quantity

Cumulative operational authority, Blast Radius, permission waste, and threat surface are formally defined, quantified, and tracked. Not estimated, measured.

timerJust-in-Time, Per-Action

Every action is evaluated individually against current risk. Permissions are ephemeral, time-bounded, and automatically revoked. No standing access.

descriptionExplainable Decisions

Every grant or denial is accompanied by a causal explanation citing specific policy rules, risk scores, and contributing factors. Decisions are reproducible and auditable.

Example Applications

Agent fleets in financial services. A bank deploys dozens of AI agents across trading, compliance, customer service, and risk management. Each agent has a registered profile with approved boundaries. The governance engine makes per-action decisions based on current risk, ensuring that a customer service agent cannot access trading systems even if a delegation chain would technically permit it.

Multi-agent workflows. A complex process involves multiple agents collaborating: one gathers data, another analyses it, a third generates recommendations. The system tracks authority accumulation across the chain, measuring the combined Blast Radius and enforcing that no agent in the chain exceeds its approved scope.

CI/CD deployment governance. Before a new version of an agent is deployed, the governance gate compares its capabilities against the approved baseline. If the new version requests access to data domains or tools not in its profile, deployment is blocked until a steward reviews and approves the change.

Edge computing governance. Agents deployed at edge locations (branches, factories, IoT environments) operate with local enforcement points that cache policy decisions. When connectivity to the central registry is impaired, edge nodes apply deny-by-default rules for novel or high-risk requests.

The Ten-Layer Agent Profile

Every agent in the registry is described by a structured profile organised into ten layers across four groups. This profile is the single source of truth for what the agent is, what it can access, what policies govern it, and why it exists.

Layer Name What It Captures
GROUP A: IDENTITY, AUTHORITY, AND DELEGATION
1Agent Identity and ProvenanceUnique identifier, Digital Birth Certificate, ownership, version history, cryptographic attestation, lifecycle state
2Authority and Invocation ContextWho or what can invoke the agent, roles and trust levels of invokers, invocation channel constraints
3Delegated AuthorityDelegation chains, inherited permissions from upstream entities, delegation history, basis for Cumulative Operational Authority calculation
GROUP B: RESOURCES, ENVIRONMENT, AND CAPABILITIES
4Data DomainsDatabases, file shares, APIs, data classifications, sensitivity levels, access modes (read/write/delete/export)
5Environment and Execution ContextRuntime environments (production, staging, development), compute context, network zones, associated risk profiles
6Tools, Services, and ChannelsSpecific APIs, tools, communication channels, connector versions, usage patterns
7Inherent Agent CapabilitiesFunctional capabilities: code execution, file system access, network communication, tool-use primitives
GROUP C: POLICY AND RISK
8Policy and Access Control RulesDeclarative policy rules (Compliance-as-Code), enforcement mechanisms, policy versions
9Risk Profiles and CategorisationBlast Radius variants, operational waste scores, Multi-Dimensional Risk Vector, risk classification
GROUP D: MISSION AND INTENT
10Mission and IntentMission statement, authorised use cases, task adherence boundaries, purpose measurement criteria, Commander's Intent anchor

The profile is versioned and cryptographically signed. Every change creates a new version with full provenance. The runtime enforcement engine uses this profile as a non-bypassable constraint for every privilege decision.

Mission Planning and Concept of Operations

The patent explicitly frames agent governance through the lens of military Concept of Operations (ConOps) doctrine. The same principles that govern how military units plan, authorise, and control operations in uncertain environments apply directly to autonomous AI agents operating in complex enterprise systems.

Commander's Intent (The "Why")

In military doctrine, Commander's Intent conveys the strategic objective and acceptable end state, enabling units to act under uncertainty while remaining aligned to the mission. In this system, Commander's Intent is represented in Layer 10 as the agent's mission and intent. It defines why the agent exists, what outcomes are acceptable, and what boundaries must not be crossed. Real-time decisions reference this intent to determine whether a requested action advances the mission within the risk tolerances defined by registry policy.

Tasks and Directives (The "What")

Permitted tasks describe the discrete activities the agent is allowed to perform: preparing a report, querying a database, generating a recommendation. Operational constraints articulate the limits that govern those activities: read-only access, no export of personal data, defined resource or scope boundaries. During evaluation, the runtime engine matches a privilege request to the relevant permitted task and applies the applicable constraints to determine whether to grant, downgrade, sandbox, or deny the request.

Mission Command and Disciplined Initiative (The "How")

The overall system embodies the philosophy of Mission Command. It provides the Commander's Intent and specific Tasks and Directives via the registry. The agent is then empowered to exercise disciplined initiative: formulating its own plan of action to achieve its goals. The system's real-time controls and pre-deployment gates ensure this initiative remains within the governed bounds defined by its directives and overall intent.

Lean Six Sigma in Agent Governance

The system applies Lean Six Sigma principles to AI agent management, treating governance failures as measurable defects rather than subjective compliance gaps. The real-time engine functions as a tactical command-and-control system that operationalises the DMAIC cycle against the versioned baseline maintained in the Agent Registry.

DMAIC as Tactical Command and Control

Define: The Registry provides the signed, authoritative baseline for each agent: identity bindings, approved capabilities, data domains, delegation boundaries, policy thresholds, and Layer 10 mission and intent. This baseline constitutes the Commander's Intent and rules of engagement that the runtime plane consumes prior to any action.

Measure: Upon a privilege request, the system validates identity against the registry record and computes a Dynamic Blast Radius and Multi-Dimensional Risk Vector. Runtime telemetry and decision outcomes are captured as observed operational reality.

Analyse: The Policy Decision Point evaluates measured risk against version-controlled policy thresholds sourced from the registry and issues a decision with an accompanying explanation that records the causal path and policy rule version.

Improve: Feedback from runtime decisions and telemetry drives design-time improvement: developers and stewards refine capabilities, scopes, delegations, and policies in the registry. The CI/CD governance gate enforces improved configurations by blocking non-conformant builds.

Control: Inline control is exercised by the Provisioning Orchestrator, which mints minimal, time-boxed credentials consistent with the registry baseline and revokes them on completion, timeout, risk escalation, or publication of new signed policy versions.

Operational Waste (Muda)

Borrowed from lean manufacturing, the system quantifies five categories of governance waste for every agent:

Waste TypeWhat It Measures
Permission WasteAuthority held beyond what the agent's declared mission requires. The gap between what an agent can do and what it needs to do.
Capability WasteInherent capabilities (code execution, network access, file system write) that exceed the mission's requirements and create unnecessary risk.
Exposure WasteAccess to data domains, classifications, or systems beyond the agent's operational need.
Transport WasteUnnecessary network hops, environment transitions, or cross-system calls that introduce latency and expand the attack surface.
Defect WasteHistorical error rates, policy violations, and revocation events that indicate process fragility.

Error-Proofing (Poka-Yoke)

Two types of error-proofing mechanism from manufacturing quality control are embedded in the architecture:

  • Control-type (CI/CD Governance Gate): A hard stop that intercepts deployments, compares new artefacts against the approved baseline, recalculates risk metrics, and halts non-compliant builds before they reach production.
  • Warning-type (Risk Simulation Engine): A "what-if" analysis tool that lets developers project the governance impact of proposed changes before committing code, surfacing issues during development rather than after deployment.

Why Current Approaches Fall Short

Existing security and governance technologies were designed for human users or deterministic software. None addresses the unique challenge of governing non-deterministic autonomous agents with cumulative operational authority.

Technology / Approach What It Does Gap
RBAC / ABAC Static role/attribute entitlements Cannot model dynamic delegated authority for non-deterministic agents.
OPA / Policy Engines Enforce rules when supplied attributes Lack a governance source of truth. Cannot detect permission waste.
IAM / PKI / SPIFFE Issue and rotate identities and secrets Do not model dynamic authority accumulation. Permission paradox at scale.
PAM / Just-in-Time Access Broker credentials, time-limit roles Do not evaluate full contextual chain from caller through agent to target.
Zero Trust Session establishment and micro-segmentation Once connected, per-action authorisation left to the application.
SIEM / SOAR Reactive post-hoc detection Cannot deliver preventative per-action authorisation at millisecond scale.
Runtime Governance Right-of-boom monitoring No left-of-boom design-time governance. No system of record before deployment.

Key Concepts

The core terminology of Agent Lifecycle Governance and the Unified Governance architecture.

Cumulative Operational Authority
The total real authority an agent possesses at any moment, including explicitly granted permissions, inherited permissions from delegation chains, and accumulated access across systems. Invisible to conventional IAM tools.
Blast Radius
A family of metrics measuring the total potential impact of an agent. Five variants are defined:
  • Static: Design-time estimate based on the approved profile.
  • Dynamic: Real-time calculation incorporating current context and delegation state.
  • Simulated: Future-state projection for pre-deployment impact assessment.
  • Aggregate: Combined exposure for teams of collaborating agents.
  • Residual: Remaining exposure after mitigating controls are applied.
Operational Waste (Muda)
Excess, un-needed authority that an agent holds beyond what is required for its declared mission. Analogous to lean manufacturing waste. Quantified and tracked as a governance metric.
Mission
The approved purpose and scope of an agent, defined in the Agent Registry. All runtime decisions are evaluated against the declared mission.
Intent
The inferred purpose of a specific agent action at runtime. Compared against the declared mission to detect drift or scope creep.
Threat Surface
The total attack surface of an agent, encompassing authorised access, delegated authority, data domains, environment bindings, tool integrations, and capability grants.
Agent Trust Score
A composite metric quantifying the overall governance posture of an agent on a 0 to 1000 scale. Higher scores indicate lower risk and tighter governance alignment.
Multi-Dimensional Risk Vector
A structured risk assessment across multiple orthogonal dimensions (authority, data, environment, tools, capabilities, delegation). Enables granular, per-dimension risk thresholds rather than a single aggregate score.
Ten-Layer Governance Data Model
The structured profile schema for each agent, comprising ten layers: identity, authority, delegation chains, data access, environment context, tools, capabilities, policy rules, risk metrics, and declared purpose.
Digital Birth Certificate
A verifiable cryptographic identity issued to every agent upon registration. Based on industry-standard protocols and coupled with the agent's structured governance profile.

Scoring

The Agent Trust Score provides a single metric for the governance posture of any agent. Higher scores indicate lower risk.

AgentTrustScore = 1000 × (1 - Riskraw)
  • Riskraw: The normalised aggregate risk score (0.0 to 1.0) derived from the Multi-Dimensional Risk Vector.
  • A score of 1000 indicates zero measured risk. A score of 0 indicates maximum risk exposure.

The Threat Surface is the union of all access domains available to an agent.

TSagent = Aauth ∪ Adel ∪ Ddata ∪ Eenv ∪ Ttools ∪ Ccap
  • Aauth: Directly authorised access.
  • Adel: Delegated authority from upstream callers.
  • Ddata: Data domains the agent can reach.
  • Eenv: Environment bindings (networks, infrastructure).
  • Ttools: External tool integrations.
  • Ccap: Capability grants (model access, API permissions).

Regulatory Alignment

The system addresses risk management, access control, logging, transparency, and human oversight requirements across major regulatory frameworks. Its Compliance-as-Code engine allows regulatory requirements to be expressed as declarative, version-controlled policy rules that are automatically enforced and audited.

gavelEU Artificial Intelligence Act

Human oversight is delivered through an interactive oversight mechanism that pauses execution when risk falls within a mandatory-review range, presenting stewards with synthesised explanations, Blast Radius visualisation, and approve, deny, or modify controls. Transparency is achieved via cryptographically bound audit trails recording causal keys, policy versions, and affected targets. Risk management relies on quantitative scoring using Blast Radius, Operational Waste, and Multi-Dimensional Risk Vector for continuous rather than periodic assessment.

View guide arrow_forward
checklistNIST AI Risk Management Framework

The ten-layer Agent Registry serves as the centralised system of record, mapping directly to the four core functions. Govern is implemented through Compliance-as-Code policy rules. Map is realised via the ten-layer data model capturing identity, authority, delegation, data domains, environment, tools, capabilities, policy, risk, and mission. Measure is delivered through quantitative metrics including Blast Radius variants, Operational Waste, and Agent Trust Score. Manage is enforced at runtime with graduated outcomes for each privilege decision.

View guide arrow_forward
shield_lockZero Trust Architecture (NIST SP 800-207)

Every agent action undergoes per-action, just-in-time privilege evaluation against dynamic Blast Radius, active risk signals, and governing policy. The system enforces zero standing privileges by design: the Provisioning Orchestrator mints minimal, time-bounded credentials for each approved action, and the revocation controller automatically expires them on completion, timeout, or risk escalation.

View guide arrow_forward
workspace_premiumISO/IEC 42001

The Agent Registry functions as the central governance artefact required by the standard. Signed baselines, versioned policy, and cryptographically chained snapshots provide certification evidence. Bidirectional reconciliation between design-time profiles and runtime telemetry implements the continuous improvement cycle the standard demands.

View guide arrow_forward
account_balanceSector-Specific Requirements

Per-action audit trails, provenance-bound explanations, and quantitative risk metrics address the supervisory expectations of financial services, healthcare, and critical infrastructure regulators. Compliance-as-Code enables continuous compliance, encoding sector-specific requirements as declarative, version-controlled policy rules that are automatically enforced and audited at every privilege decision.

Learn How This Applies to Your Organisation

Schedule a complimentary briefing to discuss how agent lifecycle governance can help your institution control the authority and impact of autonomous AI agents.

Schedule a Briefing Contact Us